Argus: A General-Purpose Agentic Runtime for Long-Horizon Reasoning

Analysis

Why This Paper Matters

Long-horizon reasoning remains a critical challenge for AI agents, as tasks often require sustained effort, adaptation to new information, and recovery from failures. Traditional approaches either rely on fine-tuning models for specific tasks or use simple prompting strategies that lack persistence. Argus introduces a general-purpose agentic runtime that addresses these limitations by providing a structured environment where multiple specialized roles (Manager, Planner, Engineer, Reviewer) collaborate over durable project state. This design allows the system to persist when evidence supports the current approach and pivot when measurements reveal issues, making it highly relevant for real-world applications like software engineering and research.

The paper's emphasis on self-evolution without changing model weights is particularly significant. By evolving the runtime state and control policy, Argus can accumulate verified approaches and improve over time, offering a cost-effective alternative to retraining. This aligns with the growing interest in agentic systems that can learn from experience and adapt to new tasks, potentially reducing the need for massive compute resources.

Technical Contributions

• Multi-Agent Architecture: Argus employs four distinct roles—Manager, Planner, Engineer, and Reviewer—each with specific responsibilities, enabling modular and scalable task execution.
• Durable Project State: The runtime maintains a persistent state that captures user intent, operational objectives, constraints, and verification criteria, allowing for continuity across long-horizon tasks.
• Role-Owned Review and Verification: All changes to the runtime state, including memories, skills, procedures, and routing decisions, undergo role-owned review and task-native verification when available, ensuring quality and reliability.
• Self-Evolution via Control Policy: Argus evolves its control policy and runtime state based on accumulated experience, without updating model weights, enabling continuous improvement.
• Operator Escalation Points: The system allows for human intervention at defined points, balancing autonomy with safety and oversight.

Results

Argus achieves approximately 78% on SWE-Bench Pro, a significant improvement over the 59% baseline for Direct Copilot, while using only 1.41 times the aggregate tokens. On AARRI-Bench, it reaches 76.8%, and on mathematical data synthesis, it shows a 28.0-point gap, indicating strong performance across diverse domains. The verification-gated self-evolution leads to measurable efficiency gains: mature SWE-Bench waves use 21% fewer solve-input tokens and 15% less active workflow time per task compared to startup waves. The system also records 34 verifier recoveries and 22 strict review-loop rescues, demonstrating its ability to correct errors and recover from failures. Beyond benchmarks, Argus contributed an optimized RWKV6 kernel that was merged upstream, and a multi-day mathematics campaign retained falsified routes and proof-backed frontier updates, showcasing real-world applicability.

Significance

Argus represents a paradigm shift in agentic AI, showing that fixed-weight models can achieve high performance on long-horizon tasks through a self-evolving runtime. This has broad implications for reducing the cost and complexity of deploying AI agents, as it eliminates the need for frequent model retraining. The structured trajectories produced by Argus also provide valuable data for future supervised and reinforcement learning, potentially accelerating progress in these areas. As agentic systems become more prevalent, frameworks like Argus could become standard for building robust, adaptive AI assistants capable of handling complex, multi-step tasks in dynamic environments.